A new HTML backend that handles styled html (ignors it) as well as images.

Note: MyPy fails.
Seems to be a known issue with BeautifulSoup:
https://github.com/python/typeshed/pull/13604

- Fixed issues with handling nested lists.
- Fixed some issues with spaces between text fragments
- Change naming of image configuration from INLINE to EMBEDDED. Also renamed corresponding class.
- Introduced constat for default image width- / height.

Signed-off-by: vaaale <2428222+vaaale@users.noreply.github.com>

docling/backend/html_backend.py

@@ -7,6 +7,7 @@ from pathlib import Path
 from typing import Optional, Union
 
 import requests
+from PIL import Image, UnidentifiedImageError
 from bs4 import BeautifulSoup, NavigableString, Tag
 from docling_core.types.doc import (
     DocItemLabel,
@@ -17,8 +18,7 @@ from docling_core.types.doc import (
     TableData,
 )
 from docling_core.types.doc.document import ContentLayer, ImageRef
-from PIL import Image, UnidentifiedImageError
-from pydantic import AnyUrl, HttpUrl, ValidationError
+from pydantic import AnyUrl, ValidationError
 from typing_extensions import override
 
 from docling.backend.abstract_backend import DeclarativeDocumentBackend
@@ -27,15 +27,17 @@ from docling.datamodel.document import InputDocument
 
 _log = logging.getLogger(__name__)
 
+DEFAULT_IMAGE_WIDTH = 128
+DEFAULT_IMAGE_HEIGHT = 128
+
 # Tags that initiate distinct Docling items
 _BLOCK_TAGS = {"h1", "h2", "h3", "h4", "h5", "h6", "p", "ul", "ol", "table"}
 
 
 class ImageOptions(str, Enum):
     """Image options for HTML backend."""
-
     NONE = "none"
-    INLINE = "inline"
+    EMBEDDED = "embedded"
     REFERENCED = "referenced"
 
 
@@ -49,7 +51,6 @@ class BaseHTMLDocumentBackend(DeclarativeDocumentBackend):
     ):
         super().__init__(in_doc, path_or_stream)
         self.image_options = image_options
-        self.soup: Optional[Tag] = None
         try:
             raw = (
                 path_or_stream.getvalue()
@@ -88,35 +89,27 @@ class BaseHTMLDocumentBackend(DeclarativeDocumentBackend):
             binary_hash=self.document_hash,
         )
         doc = DoclingDocument(name=self.file.stem or "file", origin=origin)
-        _log.debug("Starting HTML conversion...")
-        if not self.is_valid():
-            raise RuntimeError("Invalid HTML document.")
-        assert self.soup is not None
-
-        # Remove all script/style content
+        title = self.soup.find("title")
+        if title:
+            doc.add_title(title.get_text())
+        # remove scripts/styles
         for tag in self.soup.find_all(["script", "style"]):
             tag.decompose()
 
         body = self.soup.body or self.soup
-        # Normalize <br> tags to newline strings
+        # normalize <br>
         for br in body.find_all("br"):
             br.replace_with(NavigableString("\n"))
 
-        # Decide content layer by presence of headers
         headers = body.find(list(_BLOCK_TAGS))
         self.content_layer = (
             ContentLayer.BODY if headers is None else ContentLayer.FURNITURE
         )
 
-        # Walk the body to build the DoclingDocument
         self._walk(body, doc, parent=doc.body)
         return doc
 
     def _walk(self, element: Tag, doc: DoclingDocument, parent) -> None:
-        """
-        Recursively walk element.contents, buffering inline text across tags like <b> or <span>,
-        emitting text nodes only at block boundaries, and extracting images immediately.
-        """
         buffer: list[str] = []
 
         def flush_buffer():
@@ -126,88 +119,93 @@ class BaseHTMLDocumentBackend(DeclarativeDocumentBackend):
             buffer.clear()
             if not text:
                 return
-            # Split on newlines for <br>
             for part in text.split("\n"):
                 seg = part.strip()
                 if seg:
                     doc.add_text(DocItemLabel.TEXT, seg, parent=parent)
 
         for node in element.contents:
-            # Skip scripts/styles
             if isinstance(node, Tag) and node.name.lower() in ("script", "style"):
                 continue
-            # Immediate image extraction
             if isinstance(node, Tag) and node.name.lower() == "img":
                 flush_buffer()
                 self._emit_image(node, doc, parent)
                 continue
-            # Block-level element triggers flush + handle
             if isinstance(node, Tag) and node.name.lower() in _BLOCK_TAGS:
                 flush_buffer()
                 self._handle_block(node, doc, parent)
-            # Inline tag with nested blocks: recurse
             elif isinstance(node, Tag) and node.find(list(_BLOCK_TAGS)):
                 flush_buffer()
                 self._walk(node, doc, parent)
-            # Inline text
             elif isinstance(node, Tag):
                 buffer.append(node.get_text())
             elif isinstance(node, NavigableString):
                 buffer.append(str(node))
 
-        # Flush any remaining text
         flush_buffer()
 
     def _handle_block(self, tag: Tag, doc: DoclingDocument, parent) -> None:
         tag_name = tag.name.lower()
-        if tag_name == "h1":
-            text = tag.get_text(strip=True)
-            if text:
-                doc.add_title(text, parent=parent)
-            for img_tag in tag.find_all("img", recursive=True):
-                self._emit_image(img_tag, doc, parent)
-        elif tag_name in {"h2", "h3", "h4", "h5", "h6"}:
+
+        if tag_name in {"h1", "h2", "h3", "h4", "h5", "h6"}:
             level = int(tag_name[1])
-            text = tag.get_text(strip=True)
+            text = tag.get_text(strip=False)
             if text:
-                doc.add_heading(text, level=level, parent=parent)
+                doc.add_heading(text.strip(), level=level, parent=parent)
             for img_tag in tag.find_all("img", recursive=True):
                 self._emit_image(img_tag, doc, parent)
+
         elif tag_name == "p":
             for part in tag.get_text().split("\n"):
                 seg = part.strip()
                 if seg:
                     doc.add_text(DocItemLabel.TEXT, seg, parent=parent)
-                for img_tag in tag.find_all("img", recursive=True):
-                    self._emit_image(img_tag, doc, parent)
+            for img_tag in tag.find_all("img", recursive=True):
+                self._emit_image(img_tag, doc, parent)
+
         elif tag_name in {"ul", "ol"}:
-            is_ordered = tag_name == "ol"
-            group = (
+            is_ordered = (tag_name == "ol")
+            # Create the list container
+            list_group = (
                 doc.add_ordered_list(parent=parent)
                 if is_ordered
                 else doc.add_unordered_list(parent=parent)
             )
+
+            # For each top-level <li> in this list
             for li in tag.find_all("li", recursive=False):
-                li_text = li.get_text(separator=" ", strip=True)
+                # 1) extract only the "direct" text from this <li>
+                parts: list[str] = []
+                for child in li.contents:
+                    if isinstance(child, NavigableString):
+                        text_part = child.strip()
+                        if text_part:
+                            parts.append(text_part)
+                    elif isinstance(child, Tag) and child.name not in ("ul", "ol"):
+                        text_part = child.get_text(separator=" ", strip=True)
+                        if text_part:
+                            parts.append(text_part)
+                li_text = " ".join(parts)
+
+                # 2) add the list item
                 li_item = doc.add_list_item(
-                    text=li_text, enumerated=is_ordered, parent=group
+                    text=li_text, enumerated=is_ordered, parent=list_group
                 )
-                # Nested lists inside <li>
-                for sub in li.find_all(["ul", "ol"], recursive=False):
-                    self._handle_block(sub, doc, parent=group)
+
+                # 3) recurse into any nested lists, attaching them to this <li> item
+                for sublist in li.find_all(["ul", "ol"], recursive=False):
+                    self._handle_block(sublist, doc, parent=li_item)
+
+                # 4) extract any images under this <li>
                 for img_tag in li.find_all("img", recursive=True):
                     self._emit_image(img_tag, doc, li_item)
+
         elif tag_name == "table":
-            # Add table item and extract nested images
             data = self._parse_table(tag, doc, parent)
             doc.add_table(data=data, parent=parent)
 
     def _emit_image(self, img_tag: Tag, doc: DoclingDocument, parent) -> None:
-        """
-        Helper to create a PictureItem (with optional CAPTION) for an <img> tag.
-        """
-
-        if ImageOptions.NONE == self.image_options:
+        if self.image_options == ImageOptions.NONE:
             return
 
         alt = (img_tag.get("alt") or "").strip()
@@ -215,46 +213,40 @@ class BaseHTMLDocumentBackend(DeclarativeDocumentBackend):
         if alt:
             caption_item = doc.add_text(DocItemLabel.CAPTION, alt, parent=parent)
 
-        src_url = img_tag.get("src")
-        width = img_tag.get("width", "128")
-        height = img_tag.get("height", "128")
-        img_ref = None
-        if ImageOptions.INLINE == self.image_options:
+        src_url = img_tag.get("src", "")
+        width = img_tag.get("width", str(DEFAULT_IMAGE_WIDTH))
+        height = img_tag.get("height", str(DEFAULT_IMAGE_HEIGHT))
+        img_ref: Optional[ImageRef] = None
+
+        if self.image_options == ImageOptions.EMBEDDED:
             try:
                 if src_url.startswith("http"):
                     img = Image.open(requests.get(src_url, stream=True).raw)
-                elif src_url.startswith("file:"):
-                    img = Image.open(src_url)
                 elif src_url.startswith("data:"):
-                    image_data = re.sub("^data:image/.+;base64,", "", src_url)
-                    img = Image.open(BytesIO(base64.b64decode(image_data)))
+                    data = re.sub(r"^data:image/.+;base64,", "", src_url)
+                    img = Image.open(BytesIO(base64.b64decode(data)))
                 else:
                     return
-                img_ref = ImageRef.from_pil(img, dpi=int(img.info.get("dpi")[0]))
-            except (FileNotFoundError, UnidentifiedImageError) as ve:
-                _log.warning(f"Could not load image (src={src_url}): {ve}")
+                img_ref = ImageRef.from_pil(img, dpi=int(img.info.get("dpi", (72,))[0]))
+            except (FileNotFoundError, UnidentifiedImageError) as e:
+                _log.warning(f"Could not load image (src={src_url}): {e}")
                 return
-        elif ImageOptions.REFERENCED == self.image_options:
+
+        elif self.image_options == ImageOptions.REFERENCED:
             try:
-                img_url = AnyUrl(src_url)
                 img_ref = ImageRef(
-                    uri=img_url,
+                    uri=AnyUrl(src_url),
                     dpi=72,
                     mimetype="image/png",
                     size=Size(width=float(width), height=float(height)),
                 )
-            except ValidationError as ve:
-                _log.warning(f"Could not load image (src={src_url}): {ve}")
+            except ValidationError as e:
+                _log.warning(f"Could not load image (src={src_url}): {e}")
                 return
 
         doc.add_picture(image=img_ref, caption=caption_item, parent=parent)
 
     def _parse_table(self, table_tag: Tag, doc: DoclingDocument, parent) -> TableData:
-        """
-        Convert an HTML table into TableData, capturing cell spans and text,
-        and emitting any nested images as PictureItems.
-        """
-        # Build TableData
         rows = []
         for sec in ("thead", "tbody", "tfoot"):
             section = table_tag.find(sec)
@@ -262,9 +254,11 @@ class BaseHTMLDocumentBackend(DeclarativeDocumentBackend):
                 rows.extend(section.find_all("tr", recursive=False))
         if not rows:
             rows = table_tag.find_all("tr", recursive=False)
+
         occupied: dict[tuple[int, int], bool] = {}
         cells: list[TableCell] = []
         max_cols = 0
+
         for r, tr in enumerate(rows):
             c = 0
             for cell_tag in tr.find_all(("td", "th"), recursive=False):
@@ -292,9 +286,11 @@ class BaseHTMLDocumentBackend(DeclarativeDocumentBackend):
                         occupied[(r + dr, c + dc)] = True
                 c += cs
             max_cols = max(max_cols, c)
-        # Emit images inside this table
+
+        # emit any images in the table
         for img_tag in table_tag.find_all("img", recursive=True):
             self._emit_image(img_tag, doc, parent)
+
         return TableData(table_cells=cells, num_rows=len(rows), num_cols=max_cols)
 
 
@@ -308,14 +304,14 @@ class HTMLDocumentBackend(BaseHTMLDocumentBackend):
         super().__init__(in_doc, path_or_stream, image_options=ImageOptions.NONE)
 
 
-class HTMLDocumentBackendImagesInline(BaseHTMLDocumentBackend):
+class HTMLDocumentBackendImagesEmbedded(BaseHTMLDocumentBackend):
     @override
     def __init__(
         self,
         in_doc: InputDocument,
         path_or_stream: Union[BytesIO, Path],
     ):
-        super().__init__(in_doc, path_or_stream, image_options=ImageOptions.INLINE)
+        super().__init__(in_doc, path_or_stream, image_options=ImageOptions.EMBEDDED)
 
 
 class HTMLDocumentBackendImagesReferenced(BaseHTMLDocumentBackend):